Custom machine learning architectures: towards realtime anomaly detection for flight testing

Abstract

Test flight of a new commercial aeroplane is crucial in validating the functionality, safety and performance of the new aeroplane design before its batch manufacturing can take place. Massive amounts of data streams are typically generated from thousands of sensors on an aeroplane during test flight, which require realtime processing to detect anomaly and to predict malfunctions for emergency response. This paper provides an overview of recent research in custom machine learning architectures which have shown promise for highspeed data processing, and proposes a time series learning model based on LSTM (Long Short Term Memory). This LSTM model is adopted for realtime data analysis used in anomaly detection for the COMAC C919 test flight. A custom architecture targeting FPGA (Field Programmable Gate Array) implementation for the proposed approach can be embedded into realtime data analysis and processing platforms for large commercial aircraft.